clustra: A multi-platform k-means clustering algorithm for analysis of longitudinal trajectories in large electronic health records data

TL;DR

This paper introduces clustra, a multi-platform k-means clustering algorithm designed to analyze longitudinal trajectories in large electronic health records, enabling better understanding of patient data over time.

Contribution

The paper presents a novel multi-platform implementation of a k-means clustering algorithm with splines for longitudinal data, compatible with R and SAS, and includes diagnostic tools.

Findings

Comparable results achieved with R and SAS implementations

Effective clustering of blood pressure trajectories in EHR data

Supports large-scale longitudinal data analysis

Abstract

Background and Objective: Variables collected over time, or longitudinally, such as biologic measurements in electronic health records data, are not simple to summarize with a single time-point, and thus can be more holistically conceptualized as trajectories over time. Cluster analysis with longitudinal data further allows for clinical representation of groups of subjects with similar trajectories and identification of unique characteristics, or phenotypes, that can be investigated as risk factors or disease outcomes. Some of the challenges in estimating these clustered trajectories lie in the handling of observations at inconsistent time intervals and the usability of algorithms across programming languages. Methods: We propose longitudinal trajectory clustering using a k-means algorithm with thin-plate regression splines, implemented across multiple platforms, the R package clustra and corresponding \SAS macros. The \SAS macros accommodate flexible clustering approaches, and also include visualization of the clusters, and silhouette plots for diagnostic evaluation of the appropriate cluster number. The R package, designed in parallel, has similar functionality, with additional multi-core processing and Rand-index-based diagnostics. Results: The package and macros achieve comparable results when applied to an example of simulated blood pressure measurements based on real data from Veterans Affairs Healthcare recipients who were initiated on anti-hypertensive medication. Conclusion: The R package clustra and the SAS macros integrate a K-means clustering algorithm for longitudinal trajectories that operates with large electronic health record data. The implementations provide comparable results in both platforms, satisfying the needs of investigators familiar with, or constrained by access to, one or the other platform.